结合机器学习模型的早产儿ICU死亡风险评估与可解释性分析

Objective: Aimed at using machine learning algorithms to predict the risk of neonatal ICU mortality, providing clinicians with an early diagnosis and risk assessment tool to assist in decision-making. Methods: Retrospectively collecting clinical data of preterm infants from the PIC database. Cases were divided into mortality and survival groups based on ICU outcomes. Key clinical characteristics potentially affecting preterm infant outcomes were screened using LASSO regression analysis and multivariate logistic regression analysis. The study balanced the data using the SMOTE algorithm and constructed predictive models using seven machine learning models (e.g., LightGBM, random forest), evaluating their performance. Model interpretation was performed using the Shapley Additive Explanations (SHAP) algorithm. Results: A total of 923 infants were included in the final analysis. The survival group comprised 886 infants, and the death group comprised 37 infants. A total of 38 clinical characteristics were collected. LASSO screening identified 8 variables significantly associated with neonatal ICU mortality, including lactate, chloride concentration, neutrophils, and red blood cell distribution width. Multivariate logistic regression analysis revealed that lactate and respiratory rate were independent predictors of neonatal ICU outcomes. The LightGBM model achieved an AUC of 0.972 and outperformed other models in terms of accuracy and precision. Furthermore, SHAP analysis enhanced model interpretability. The results indicated that respiratory rate and lactate contributed most significantly to the prediction of infant mortality risk. Conclusion: This study provides reliable tools for early identification and intervention of preterm infant outcomes, emphasizing the importance of key physiological indicators. Future multi-center data validation is needed to enhance the models generalizability and further optimize algorithm performance.